Agent Mode
An autonomous AI coding assistant that uses tools to read files, write code, run tests, and manage your project - all while keeping you in control.
Getting Started
Switching to Agent Mode
- Open your workspace in the Generation page
- Look for the Mode Toggle in the header
- Click "Agent" to switch from Classic mode
- You'll see a green "Tool-based agent" indicator when active
Your First Conversation
Try simple requests like:
- •"Read the package.json file"
- •"List all TypeScript files"
- •"Create a new README.md file"
How Agent Mode Works
Available Tools (32 Total)
File Operations (8 tools)
- •Read files
- •Write new files
- •Edit existing files
- •Search file contents
- •List files by pattern
- •Delete files (with confirmation)
- •Apply code diffs
- •Get file with dependencies
Code Execution (5 tools)
- •Run shell commands
- •Install packages (npm, pip, etc.)
- •Run tests
- •Lint code
- •Format code
Git Operations (6 tools)
- •Git status and diff
- •Create commits
- •Push to remote (with confirmation)
- •Pull from remote
- •Manage branches
Context Analysis (4 tools)
- •Analyze codebase structure
- •Get component information
- •Find references
- •Trace dependencies
Web Operations (3 tools)
- •Search the web
- •Scrape websites
- •Fetch documentation
Planning (6 tools)
- •Break down complex tasks
- •Verify step completion
- •Ask clarifying questions
- •Quality validation (linting)
- •Finish task
Understanding Confirmations
The agent uses smart confirmations - you'll only be prompted for high-risk operations:
Always Confirm
(High Risk)- •Deleting files
- •Git push to remote
- •Destructive commands (rm -rf, sudo, etc.)
Sometimes Confirm
(Medium Risk)- •Modifying important files (package.json, tsconfig.json)
- •Installing new packages
- •Large refactors (5+ files)
- •Auto-fixing lint errors
- •Formatting all files
Never Confirm
(Low Risk)- •Reading files
- •Searching code
- •Git status/diff
- •Context analysis
Tips for Best Results
Be Specific
X "Fix the app"
O "Fix the authentication error when users try to login"
Provide Context
If the agent needs more information, it will ask questions like: "Which authentication method should I use?" or "Should I create tests for this component?"
Let the Agent Work
The agent operates autonomously through multiple iterations. You'll see state changes, tool executions, and progress updates. Don't interrupt - watch the magic happen!
Keyboard Shortcuts
| Shortcut | Action |
|---|---|
Cmd/Ctrl + K | Focus input field |
Cmd/Ctrl + Enter | Send message |
Cmd/Ctrl + Shift + C | Clear conversation |
Cmd/Ctrl + M | Toggle Classic/Agent mode |
Troubleshooting
Agent is Stuck
Check the current state indicator, wait for the iteration to complete, try rephrasing your request, or switch to Classic mode temporarily.
Unexpected Results
Provide more specific instructions, use follow-up messages to refine, or ask the agent to undo or modify its changes.
Error Recovery
Retryable errors (network, timeout) auto-retry up to 3 times. Non-retryable errors trigger alternative approaches. Critical errors are caught by error boundary with recovery options.